☕️ OpenAI missed its own revenue and user growth targets

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

📉 OpenAI missed its own revenue and user growth targets

✉️ Google employees urge Pichai to reject Pentagon AI deal

🎥 Google is testing AI chatbot search for YouTube

📱 Apple plans iPhone Ultra and MacBook Ultra

🇪🇺 EU orders Google to open Android to AI rivals

Plus: 🎁 14 other news you might like, 🧰 6 tools, and 📚 4 papers.

📉 OpenAI missed its own revenue and user growth targets LINK

OpenAI fell short of its internal goals for ChatGPT users and revenue last year, never reaching its target of one billion weekly active users, while CFO Sarah Friar warned that rising compute costs could outrun incoming revenue.

The company has committed roughly $600 billion to future data-center spending under Altman's bet on compute scarcity, and board directors are now questioning why he keeps chasing more computing capacity despite the slowdown.

Rival Anthropic has quietly passed OpenAI on Forge Global, trading at about $1 trillion versus OpenAI's $880 billion, and Myriad users give Anthropic a 64% chance of carrying out its IPO first.

✉️ Google employees urge Pichai to reject Pentagon AI deal LINK

Roughly 600 Google workers have signed an open letter asking CEO Sundar Pichai to walk away from talks with the Pentagon that would let the Department of Defense use the company's Gemini AI models in classified settings.

The signatories argue that contract wording is not enough protection, pointing to how Anthropic was labeled a "supply chain risk" after refusing "all lawful purposes" language, while OpenAI revised its Pentagon deal to block mass surveillance of U.S. persons.

The letter follows Google's recent rewrite of its AI Principles, which in 2018 promised staff the company would not design or deploy AI for weapons or surveillance, language that employees say has since shifted.

🎥 Google is testing AI chatbot search for YouTube LINK

Google is trying out an AI Mode-style conversational search for YouTube, and the experiment is open now to YouTube Premium subscribers in the US who are 18 or older, with plans to expand it to other users.

An "Ask YouTube" button in the search bar brings up a page that mixes summary text, bulleted milestones, timestamped longform videos, Shorts galleries, and suggested follow-up prompts related to what you searched for.

In a test about Valve's new Steam Controller, Ask YouTube got the basics right but incorrectly said the old Steam Controller had no joysticks, a reminder that these AI-built result pages can include factual errors.

📱 Apple plans iPhone Ultra and MacBook Ultra LINK

Apple is preparing to expand its Ultra branding into new product tiers, with plans for a foldable iPhone Ultra and a touchscreen OLED MacBook Ultra that will sit above the existing Pro lineup through 2027.

According to a Macworld report citing a source familiar with Apple's plans, the Ultra name gives the company a place for new form factors without disrupting the Pro lineup, which has covered iPhone, iPad, and Mac for years.

The Ultra tier lets Apple ship foldable displays and OLED touchscreen Macs in limited quantities at higher prices, fitting its shift toward raising revenue per user as global smartphone growth slows.

🇪🇺 EU orders Google to open Android to AI rivals LINK

The European Commission has told Google it must open up Android so rival AI services can match what Gemini does on the phones, a decision that came out of a specification proceeding started in January.

Gemini currently gets special treatment at the system level on any Google-powered Android phone, and the commission says too many Android experiences only work with Google's AI, which must change.

The order comes from the Digital Markets Act, which labels seven dominant firms as "gatekeepers," and the commission may force Google to make the Android AI changes this summer despite Google calling it "unwarranted intervention."

Other news you might like

• Taylor Swift files to trademark her voice, likeness to ward off AI deepfakesLINK

• Apple introduces a cheaper option for App Store subscriptionsLINK

• Meta Prepares to Undo $2.5 Billion Manus Deal After China BanLINK

• Australia to charge Big Tech companies 2% levy unless they strike local news dealsLINK

• Former Google DeepMind researcher's AI startup raises record $1.1 billion seed funding to pursue superintelligenceLINK

• Humanoid robots to become baggage handlers in Japan airport experimentLINK

• Replit's CEO says it's dumb to study computer science thinking you can make a 'boatload' at GoogleLINK

• Claude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes rogueLINK

• Intel Is Selling Duff CPUs to AI Builders as They Scramble for ChipsLINK

• Start with the sensors, then design the rest: How Zoox built its robotaxiLINK

• Introducing talkie: a 13B vintage language model from 1930LINK

• Images of Samsung's rumored smart glasses have leakedLINK

🧰 Trending tools

SureThing.io: turn GitHub skills into a coordinated AI executive team with shared memory, so business goals get hit without manual debugging.LINK

Clera: an AI talent agent that learns your job preferences via iMessage or WhatsApp and sends curated role matches with direct company introductions.LINK

Social Fetch: one API to scrape profiles, posts, comments, and video transcripts across TikTok, Instagram, YouTube, X, LinkedIn, and Facebook without rate limits.LINK

OrcaSheets AI Reports: locally-run analytics tool that lets you query unlimited data in plain English, automate reports, and avoid cloud costs.LINK

Lovable mobile app: lets you build and ship full-stack applications using AI, significantly reducing development time without writing code manually.LINK

MaxHermes by Minimax: a cloud sandbox AI agent that extracts reusable skills from completed tasks, compounding its usefulness across sessions for knowledge workers and enterprise teams.LINK

📚 Trending papers & reports

AI document delegation fails badly, as testing 19 large language models across 52 professional fields found even top models corrupt an average of 25% of document content during long editing workflows.LINK

A new Top-K algorithm for AI inference on NVIDIA Blackwell hardware achieves an average 1.88x speedup over the standard method, cutting response latency by up to 7.52% at 100K context length.LINK

AI hiring tools strongly favor resumes they wrote themselves, giving candidates who use the same AI as the screener a 23% to 60% better chance of being shortlisted than equally qualified people who wrote their own resumes.LINK

Agentic coding performance jumps significantly when an AI summarizes its own failed attempts into compact notes, helping it learn from mistakes, with Claude-4.5-Opus improving from 70.9% to 77.6% on one benchmark and 46.9% to 59.1% on another.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!